ether0

ether0 is a 24B open-weights scientific reasoning model developed by FutureHouse, specifically trained on chemistry tasks with a particular strength in designing drug-like molecules. It accepts questions in natural language, reasons through problems in natural language, and returns answers in the form of molecules. Designed to complement FutureHouse's broader suite of scientific agents — which cover literature search, hypothesis generation, and data analysis — ether0 addresses the need for specialized scientific intelligence at the model level, going beyond what general-purpose frontier models can offer for molecular work.

ether0 is built on the Mistral 24B Instruct model and trained using reinforcement learning combined with successive rounds of fine-tuning. It is available for download on Hugging Face under a permissive Apache 2.0 license, accessible as a tool within the FutureHouse platform, and integrated into FutureHouse's chemistry agent, Phoenix, for molecular design tasks.

How ether0 Works: Reinforcement Learning from the Physical World

• ether0 is a reasoning model that generates specially marked "reasoning tokens" in natural language before producing a final answer, with those tokens emerging through reinforcement learning based solely on the correctness of the final answer rather than imitation of a reference.
• Training began from the Mistral 24B Instruct model and proceeded through successive rounds of reinforcement learning and fine-tuning, combining worked-out chains of thought with correct answers to distill knowledge across models.
• A key architectural insight is the use of multiple separate specialist models whose answers and chains of thought are aggregated to train a single generalist model.
• The model was trained on approximately 50,000 examples per task — a manageable scale that FutureHouse believes is replicable across other scientific domains.
• ether0 demonstrates that small open-source models can exceed the performance of frontier models on molecular design tasks, and in some cases learn those tasks more efficiently than specialized models trained from scratch on the same data.
• The training process revealed that completely novel tasks not present in pretraining data — such as a custom "functional group" task — can be learned during the post-training reinforcement learning phase, contributing a meaningful data point to ongoing debates about when and how language models acquire reasoning capabilities.

Capabilities and Strengths

• Excels at designing drug-like molecules given constraints expressed in natural language, including generating molecules that match specified molecular formulas with correct atom counts and plausible structures.
• Outperforms frontier models such as OpenAI o3 and Claude Opus 4 on molecular design tasks despite having a significantly smaller parameter count.
• Leverages the extensive latent chemistry knowledge already stored in large language model weights, amplified through a small but targeted reinforcement learning process.
• Develops emergent reasoning behaviors during training, including mixing languages, inventing new domain-specific vocabulary (e.g., the word "reductamol"), and forming mental shortcuts in its reasoning chains.

Known Limitations

• Described as a research prototype rather than a production-ready tool; it is not optimized for general chemistry knowledge recall.
• Struggles with knowledge-based questions such as common names for compounds or obscure reaction classes.
• Does not account for molecular conformations or three-dimensional shapes.
• Does not perform well on standardized chemistry benchmarks such as ChemBench.
• Is not designed for conversational interaction; it is best suited to questions whose answers are molecules.
• Reasoning chains become increasingly alien and less interpretable the longer training continues.

Safety and Dual-Use Mitigation

• FutureHouse conducted post-training safety mitigation specifically to reduce the risk of misuse of the open-weights model.
• The model is designed to refuse requests to design internationally controlled compounds.
• The model has little to no utility for "tacit" procedural knowledge such as setting up reaction conditions or workups.
• Red-teaming of the open-weights model was conducted and reported in the accompanying manuscript.
• Deployed instances of the model on the FutureHouse platform are subject to ongoing monitoring and filtering.

Open-Source Contributions and Resources

• Model weights are released on Hugging Face under the Apache 2.0 license.
• A preprint describing the full training process, including data composition, is available on arXiv (arxiv.org/abs/2506.17238).
• The reward model and benchmark for comparing future models are publicly released, with reward code and molecular design utilities available at github.com/Future-House/ether0.
• A new benchmark dataset is available at huggingface.co/datasets/futurehouse/ether0-benchmark.
• An interactive demo is available at ether0.platform.futurehouse.org.
• Training was conducted with GPU infrastructure provided by VoltagePark, with technical support from NVIDIA; early experiments leveraged Hugging Face's TRL and OpenR1 repositories.

ether0 represents FutureHouse's early step toward building AI systems capable of making scientific connections beyond human reach. The team views the training methodology as a template that can be extended across many other scientific domains, with the ultimate goal of accelerating autonomous scientific discovery.